AMD - Zen chitchat

[watercooled]

I wonder how much of that bugginess is down to the vanilla drivers, and how much of it is down to the presumably custom HP implementation? I only ask because TechReport never mentioned any such crashes or bugs in their review.

Also I'm not sure I understand their reference to desktop Ryzen's FMA bug and compare it to the laptop crashing in Furmark - a GPU benchmark? CPUs have erratum, whether that's Intel, AMD or anyone else, and it's not usually a big deal.

I don't disagree that AMD's drivers often need more polish, but when you're comparing against Intel's GPU drivers? Yeah... no...

IIRC Intel's drivers are the only ones that have caused me BSODs across multiple versions and rollbacks. You really can't knock AMD's driver stability when comparing to Intel.

[Xlucine]

Raven ridge has a spare graphics CU? I do like the whole naming GPU's after their compute blocks thing, it's really clear with the relative rankings

[CAT-THE-FIFTH]

I wonder how much of that bugginess is down to the vanilla drivers, and how much of it is down to the presumably custom HP implementation? I only ask because TechReport never mentioned any such crashes or bugs in their review.

Also I'm not sure I understand their reference to desktop Ryzen's FMA bug and compare it to the laptop crashing in Furmark - a GPU benchmark? CPUs have erratum, whether that's Intel, AMD or anyone else, and it's not usually a big deal.

I don't disagree that AMD's drivers often need more polish, but when you're comparing against Intel's GPU drivers? Yeah... no...

IIRC Intel's drivers are the only ones that have caused me BSODs across multiple versions and rollbacks. You really can't knock AMD's driver stability when comparing to Intel.

Well as you might realise many of these reviewers seem to conveniently ignore the Intel crappy IGP drivers,so it does make me wonder whether how much of this is down to review sites being worried Intel won't supply them with review laptops - even TR got their own Intel laptops from Intel. They all seem to be trying to push it towards systems with Nvidia dGPUs.

As SJ mentioned earlier,maybe AMD really needs to start seeding units with the equivalent Intel and AMD X360 laptops to reviewers.

[CAT-THE-FIFTH]

AMD apparently working on using GDDR6 for its graphics cards:

https://translate.google.co.uk/trans...cherinterfaces

[cascadde]

HBM didn't seem to work out for them so well, even when it seems like a great technology.

[DanceswithUnix]

HBM didn't seem to work out for them so well, even when it seems like a great technology.

HBM is still a high end technology, mid range cards didn't need it and with GDDR6 still won't for some time.

[badass]

HBM is still a high end technology, mid range cards didn't need it and with GDDR6 still won't for some time.

Exactly. It looks like Vega is actually held beck by its bandwidth to the HBM due to AMD only using 2 packages with half the pins of Fury (and less than double the bandwidth/pin. Nvidias high end uses double the number of packages on something that costs an order of magnitude more.

[scaryjim]

... less than double the bandwidth/pin. ...

I suspect that's the bit that's hurting them. All the build up was about HBM2 doubling data rates, and I imagine they decided early that at 2Gbps they could get away with only 2 stacks to keep costs down. Why they've ended up stuck at 1.6Gbps I don't know, but that's sucked 20% of their projected bandwidth away ... that's got to hurt performance...

[DanceswithUnix]

Exactly. It looks like Vega is actually held beck by its bandwidth to the HBM due to AMD only using 2 packages with half the pins of Fury (and less than double the bandwidth/pin. Nvidias high end uses double the number of packages on something that costs an order of magnitude more.

How is Vega held back by HBM? The Vega 56 and 64 have the same memory bandwidth and yet the 56 is slower so at the very least the 56 isn't bandwidth limited. They also have the same bandwidth as a 1080ti. Frankly I think Vega just needs more shaders on a bigger chip, but wonder if AMD have a design limit that holds them down to 4096 as that is what both Vega and Fury had.

[scaryjim]

How is Vega held back by HBM? The Vega 56 and 64 have the same memory bandwidth and yet the 56 is slower so at the very least the 56 isn't bandwidth limited. ....

Vega 64 stock memory clock is almost 20% higher than Vega 56. Bandwidth is 484GB/s vs 410GB/s.

Memory overclocking gains on Vega 56 aren't huge, but they are there. Gamersnexus saw gains of 3.6% from overclocking just the memory* - but that's compared to a memory overclock of ~ 19%. I haven't found a review of Vega 64 with a memory-only overclock yet, but AFAICT it doesn't go much beyond 1000MHz, so gains are going to be minimal anyway with Vega 64 already running the memory at 945MHz stock...


* Their testing didn't actually cover *just* a memory overclock: they first upped the power target by 50%, which yielded a 12.6% performance increase, then upped the memory from 800MHz to 950MHz, which gave them a further 3.6%.

[Xlucine]

Frankly I think Vega just needs more shaders on a bigger chip, but wonder if AMD have a design limit that holds them down to 4096 as that is what both Vega and Fury had.

Fury was a special case, when they had a completely understood process so they could run the die right up to the reticle size. The engineers were probably feeling really chuffed about getting the performance gains that they did without pushing up the compute units - vega is already a big chip, they don't need it to be more expensive to produce

[CAT-THE-FIFTH]

Finally a review site tested both AMD and Intel versions of the HP Envy X360:

https://www.laptopmag.com/reviews/la...-envy-x360-15t

Battery Life

You'll need to carry the Envy notebooks' power cords wherever you go, as they fail to meet the already-low standards for big-screen notebook battery life. The Laptop Mag Battery Test (web surfing at 100 nits) drained the Intel model of its charge in 5 hours and 49 minutes, while the AMD model hit empty after 5 hours and 11 minutes. Those times aren't just shorter than the 6:34 mainstream notebook average, but below the 6:14 from the Inspiron 15, the 8:41 from the Spectre x360 and the 8:59 from the Yoga 720.

[DanceswithUnix]

Vega 64 stock memory clock is almost 20% higher than Vega 56. Bandwidth is 484GB/s vs 410GB/s.

I did fact check that before posting, but guess I checked with a duff review

Overclocking the ram improves the latency as well as the bandwidth, so I would always expect some performance improvements. But an 18% overclock yielding a 3.6% performance increase isn't selling me on there being a memory access wall.

Fury was a special case, when they had a completely understood process so they could run the die right up to the reticle size. The engineers were probably feeling really chuffed about getting the performance gains that they did without pushing up the compute units - vega is already a big chip, they don't need it to be more expensive to produce

I guess I'm used to seeing 560mm^2 Titan chips at 28nm and the P100 at 610mm^2, but you are right AMD are just under performing per square mm if you compare with the 1080ti.

[scaryjim]

... Overclocking the ram improves the latency as well as the bandwidth, so I would always expect some performance improvements. But an 18% overclock yielding a 3.6% performance increase isn't selling me on there being a memory access wall. ...

Good point on the latency, hadn't considered that They're certainly not up against a memory wall - and in itself that's quite impressive given they've got less bandwidth than Fury X had but significantly higher performance. Not sure if that's down to improved compression techniques or if Fury X was just giving bandwidth away for Christmas!

[CAT-THE-FIFTH]

I did fact check that before posting, but guess I checked with a duff review

Overclocking the ram improves the latency as well as the bandwidth, so I would always expect some performance improvements. But an 18% overclock yielding a 3.6% performance increase isn't selling me on there being a memory access wall.



I guess I'm used to seeing 560mm^2 Titan chips at 28nm and the P100 at 610mm^2, but you are right AMD are just under performing per square mm if you compare with the 1080ti.

For the last couple of years Nvidia has gone to distinct lines for compute and gaming -AMD hasn't and they also don't build the largest GPUs too.

Compare Vega to the GP102 at something like deep learning or other compute operations for example.

The lower gaming performance is because AMD is doing a one size fits all strategy. If they want to catch Nvidia in performance per mm2 in gaming they really need to move to dedicated gaming GPUs.

[DanceswithUnix]

Good point on the latency, hadn't considered that They're certainly not up against a memory wall - and in itself that's quite impressive given they've got less bandwidth than Fury X had but significantly higher performance. Not sure if that's down to improved compression techniques or if Fury X was just giving bandwidth away for Christmas!

Caching for the wide slow pipe that is HBM must be, well, interesting so I presume they will have built lessons learnt from Fury into the cache of Vega.

For the last couple of years Nvidia has gone to distinct lines for compute and gaming -AMD hasn't and they also don't build the largest GPUs too.

Compare Vega to the GP102 at something like deep learning or other compute operations for example.

The lower gaming performance is because AMD is doing a one size fits all strategy. If they want to catch Nvidia in performance per mm2 in gaming they really need to move to dedicated gaming GPUs.

Last couple of years? The GTX460 is some time back now, doesn't time fly!
Still, the 1080ti is a reworked Titan, not a consumer part, and whilst you can find compute tasks where the Vega is faster it generally loses.

https://www.anandtech.com/bench/product/1937?vs=1939